###### tags: `Supervision`

# Agenda for 18.06.2021 Supervision Meeting
## List of things to be discussed:
* FairSMOTEBoost
* Online Learning
### 1- FairSMOTEBoost:
Last week I contributed majority of my time on the implementation and I found it more interesting. Although, I have some questions about basic concepts (both in methodology concepts and in the implementation).
#### Questions:
* Does SMOTEBoost use the same boosting framework as the AdaBoost after synthetic data augmentation or it uses a different one? In the code I see a version called SAMME Adaboost [1] is implemented for SMOTEBoosting.
* In the classic boosting we have a stump for each attribute. I guess it is not a case here. Do we have any limitation for the number of weak learners ?
* I also confused a bit the parameter of Boosting Rounds with the iterations and number of estimators. If you could review these parameters from the AdaBoost or SMOTEBoost algorithms briefly for me.

#### Progress:
I changed parameters to see how the results differ. There are interesting results. With increasing the value of sample_percentage (you could see this parameter in the last line of function Evaluation.py) the accuracy and balanced accuracy of the FairSMOTEBoost increases. Here are some comparisons for Adult data with 1%, 5%, 10% and 20%:




The last one is smaple_percentage = 20%

Also I tried to observe the behaviour of weights by changing the boosting rounds parameter. In the following you see the performance evolution and wights of each of the methods on Banks data (performance on left and weights on right). The graphs refer to AdaBoost, FairSMOTEBost, SMOTEBOOST and RUSBoost respectively.


Now we see the similar results n Bank data for the same order of algorithms using 200 Boosting rounds. again performance on left and weights on right for ech method. The graphs refer to AdaBoost, FairSMOTEBost, SMOTEBOOST and RUSBoost respectively:




I want to have your opinion and analyze of the obtained results. Thank you
### 2- Online Learning:

What do you suggest me to read and work on in the context of online learning.

#### References
[1] J. Zhu, H. Zou, S. Rosset, T. Hastie, "Multi-class AdaBoost", 2009

(In the implementation it is cited as a 2009 paper but I searched and found a 2006 paper with this title.)